ICON Easi3R: Estimating Disentangled Motion from DUSt3R Without Training

Xingyu Chen1     Yue Chen1     Yuliang Xiu1,2     Andreas Geiger3     Anpei Chen1,3    
1Westlake University      2Max Planck Institute for Intelligent Systems      3University of Tübingen, Tübingen AI Center
arXiv Project Page ⚖️ Comparison Code

Disentangled Reconstruction

We introduce the spatial and temporal attention mechanism behind DUSt3R, enabling robust dynamic object segmentation.
This segmentation efficiently disentangles object and camera motion, enabling DUSt3R to robustly reconstruct 4D.

4D Reconstruction

Static Scene

Dynamic Object

breakdance-flare elephant car-shadow bear bus dog paragliding snowboard BruceLee1 dance6 fighting6 LaLaLand1.2

Results are downsampled for efficient online rendering

Left-click and drag to rotate
Right-click and drag or WASD to move
Scroll to zoom
Click to pause